Possibly non-reproduceable fits
Dear All, Please bear with me here - it's possible I've done something stupid without spotting it - but I was fitting some data in Artemis and have two fits, that I am convinced should be identical (same data, same k and r windows, same paths, same GDS etc. etc.), but are not. What I did was as follows (fit numbers refer to the Artemis history file in the attached project): I set up a fit to some data, simultaneously refining at Au and Ag edges to Au-Au/Au-Ag/Ag-Au/Au-Ag paths (fits 1 & 2), and then tried a different fit, this time with Ag-Cl paths as well (fits 3&4). I then fitted the original model again (fit 5 - this was slightly different to 1 & 2, but not enough for me to notice at the time). I then closed Artemis and finished for the day. Today I came back, reloaded the project file, saved it in a new location, and tried to pick up where I left off. I performed one fit where I'd set the values to def, rather than guess, by mistake (fit 6) and then corrected this when I realised what I'd done. However when I tried to re-run the model with Ag-Cl paths I was getting wildly different results to the first time (Fits 7&8). As far as I can tell these models should be identical to fits 3&4. Both datasets were chi(k) data imported directly into Artemis. OS: Win 8.1 Ent x64 Demeter: 0.9.22 pre something x64 I've attached the following: Artemis project (artemis.fpj) Logs from fits 3 and 7 for comparison (fit3.txt, fit7.txt) Artemis log file (dartemis.log). (Also possibly worth noting is that Artemis seems to have lost the file locations (in the log files and the data window), first for the Au edge data, and then for both datasets; however when I plot the data it all seems present and correct!) Any insight gratefully received! Thanks, Ian ---- Ian Godfrey PhD Student, UCL/JAIST Programme Industrial Doctorate Centre in Molecular Modelling and Materials Science, Department of Chemistry, University College London and School of Materials Science, Japan Advanced Institute of Science and Technology i.godfrey@ucl.ac.ukmailto:i.godfrey@ucl.ac.uk i.godfrey@jaist.ac.jpmailto:i.godfrey@jaist.ac.jp
Ian, I am suspicious that the quick first shell paths are not getting saved correctly to the project file. Or perhaps not getting restored correctly from the project file. I'll look into it. Since your description isn't exactly a recipe, I am going to have to do some testing to try to find the problem. Perhaps the easiest work-around would be to not use the quick first shell tool. For the Au-Au parts of the fit, using the crystal structure for the pure metal is simple. For the Au-Ag, Ag-Au, and Au-Cl parts, it should be fairly simple to make a feff.inp file. For the mixed metal parts, try replacing the entire first shell with the other metal. For Au-Cl, try this: http://www.crystallography.net/cod/1510088.html Thanks for the bug report -- I'll let you know if I figure anything out. B On 08/04/2015 12:00 AM, Godfrey, Ian wrote:
Dear All,
Please bear with me here - it's possible I've done something stupid without spotting it - but I was fitting some data in Artemis and have two fits, that I am convinced should be identical (same data, same k and r windows, same paths, same GDS etc. etc.), but are not.
What I did was as follows (fit numbers refer to the Artemis history file in the attached project):
I set up a fit to some data, simultaneously refining at Au and Ag edges to Au-Au/Au-Ag/Ag-Au/Au-Ag paths (fits 1 & 2), and then tried a different fit, this time with Ag-Cl paths as well (fits 3&4). I then fitted the original model again (fit 5 - this was slightly different to 1 & 2, but not enough for me to notice at the time). I then closed Artemis and finished for the day. Today I came back, reloaded the project file, saved it in a new location, and tried to pick up where I left off. I performed one fit where I'd set the values to def, rather than guess, by mistake (fit 6) and then corrected this when I realised what I'd done. However when I tried to re-run the model with Ag-Cl paths I was getting wildly different results to the first time (Fits 7&8). As far as I can tell these models should be identical to fits 3&4.
Both datasets were chi(k) data imported directly into Artemis.
OS: Win 8.1 Ent x64
Demeter: 0.9.22 pre something x64
I've attached the following:
Artemis project (artemis.fpj)
Logs from fits 3 and 7 for comparison (fit3.txt, fit7.txt)
Artemis log file (dartemis.log).
(Also possibly worth noting is that Artemis seems to have lost the file locations (in the log files and the data window), first for the Au edge data, and then for both datasets; however when I plot the data it all seems present and correct!)
Any insight gratefully received!
Thanks,
Ian
----
Ian Godfrey
PhD Student,
UCL/JAIST Programme
Industrial Doctorate Centre in Molecular Modelling and Materials Science,
Department of Chemistry,
University College London
and
School of Materials Science,
Japan Advanced Institute of Science and Technology
i.godfrey@ucl.ac.uk mailto:i.godfrey@ucl.ac.uk i.godfrey@jaist.ac.jp mailto:i.godfrey@jaist.ac.jp
_______________________________________________ Ifeffit mailing list Ifeffit@millenia.cars.aps.anl.gov http://millenia.cars.aps.anl.gov/mailman/listinfo/ifeffit
-- Bruce Ravel ------------------------------------ bravel@bnl.gov National Institute of Standards and Technology Synchrotron Science Group at NSLS-II Building 535A Upton NY, 11973 Homepage: http://bruceravel.github.io/home/ Software: https://github.com/bruceravel Demeter: http://bruceravel.github.io/demeter/
I'll look into it. Since your description isn't exactly a recipe, I am going to have to do some testing to try to find the problem.
Sorry. I've never encountered (or at least noticed) it before so I'm not sure exactly what it takes to reproduce it. Probably I should have taken a bit more time to investigate before reporting it.
Perhaps the easiest work-around would be to not use the quick first shell tool. For the Au-Au parts of the fit, using the crystal structure for the pure metal is simple. For the Au-Ag, Ag-Au, and Au-Cl parts, it should be fairly simple to make a feff.inp file. For the mixed metal parts, try replacing the entire first shell with the other metal. For Au-Cl, try this: http://www.crystallography.net/cod/1510088.html
Thanks for the work around. Ian
Hi Ian, I just uploaded trial installers. Links can be found at http://bruceravel.github.io/demeter/#windows I am pretty sure that the plotting issue is resolved and that you can configure the programs to use the wxt terminal everytime. I found and solved /a/ problem with fitting. Whether I found and solved /your/ is something you'll have to tell me :) I was able to replicate something quite similar to what you described. I believe that my comment about it being a problem with quick first shell paths was incorrect and that my suggested work-around was a red herring. That is, I suspect you will see exactly the same problem with my "work-around". If the thing I fixed is the same issue that you are seeing, the fit is not retroactive. That is, I cannot fix the history in an existing project file. But I think your next project file will work as you expect. Give the new installer a try and let me know if you think it is working. And thanks, as always, for the bug reports. (Although the fitting problem was a seriously tough one!) B On 08/04/2015 11:46 AM, Godfrey, Ian wrote:
I'll look into it. Since your description isn't exactly a recipe, I am going to have to do some testing to try to find the problem.
Sorry. I've never encountered (or at least noticed) it before so I'm not sure exactly what it takes to reproduce it. Probably I should have taken a bit more time to investigate before reporting it.
Perhaps the easiest work-around would be to not use the quick first shell tool. For the Au-Au parts of the fit, using the crystal structure for the pure metal is simple. For the Au-Ag, Ag-Au, and Au-Cl parts, it should be fairly simple to make a feff.inp file. For the mixed metal parts, try replacing the entire first shell with the other metal. For Au-Cl, try this: http://www.crystallography.net/cod/1510088.html
Thanks for the work around.
Ian
_______________________________________________ Ifeffit mailing list Ifeffit@millenia.cars.aps.anl.gov http://millenia.cars.aps.anl.gov/mailman/listinfo/ifeffit
-- Bruce Ravel ------------------------------------ bravel@bnl.gov National Institute of Standards and Technology Synchrotron Science Group at NSLS-II Building 535A Upton NY, 11973 Homepage: http://bruceravel.github.io/home/ Software: https://github.com/bruceravel Demeter: http://bruceravel.github.io/demeter/
I am pretty sure that the plotting issue is resolved and that you can configure the programs to use the wxt terminal everytime.
They're definitely using wxt by default now. I see some random junk in the plot window (screenshot attached), but that is easily rectified by changing the character encoding per your previous instructions.
If the thing I fixed is the same issue that you are seeing, the fit is not retroactive. That is, I cannot fix the history in an existing project file. But I think your next project file will work as you expect.
I'm still seeing differences between apparently identical fits. I still can't figure out exactly what steps are causing the problem, but the following seems to be a fairly reliable way of doing so (not every step may be necessary, though): 1) Load chi(k) data into Artemis - I've attached a datafile I'm experiencing this issue with. 2) Setup fitting windows, paths and GDS for a QFS fit to two paths. In particular, change the fitting windows from Artemis' automatic value and set S02 to consist of two parameters - eg. ampau*xauau and ampau*xauag; set the common parameter rather than guess it (this gives us something easy to change in the later steps) 3) run a fit 4) In GDS change the set value (ampau) to something else 5) run a fit 6) Save the project, close Artemis and then reopen the project (this step seems to be fairly important) 7) run a fit 8) change the set GDS parameter back to its original value 9) run a fit 10) compare the log from the final fit to the log from the first fit In the attached Artemis file I've attached compare fit 1 to fit 4: In this case the differences are relatively small, but I still don't think they should be there! This file was produced using the 0.9.23 pre-release Artemis. Some observations: 1) I haven't seen this with every data file - it seems to be related to data quality? For instance it seems to be impossible to reproduce this problem with foil data, but the low quality data I've attached is very prone to it. 2) The effect seems most pronounced with poorer models. A model that is basically correct displays only small changes (probably small enough to be lost to the uncertainty), while one that is basically wrong may change wildly. (2b) It may be possible that the some of the smaller changes are the result of the fitting algorithm getting confused by noise and stuck in a different local min/max, rather than an actual bug, if it uses some kind of monte carlo or other random-chance based approach?) Attachments: Screenshot of plot window Example chi(k) datafile which displays this problem: ig060.cho Example fitting project: singleedgetest.fpj Thanks, Ian
participants (2)
-
Bruce Ravel -
Godfrey, Ian